Punjabi Gurmukhi input appears to mishandle addak and nasalization markers

[selsie]

Self Checks

• This template is only for bug reports. For questions, please visit [Discussions](https://github.com/fishaudio/fish-speech/discussions).
• I have thoroughly reviewed the project documentation (installation, training, inference) but couldn't find information to solve my problem. [English](https://speech.fish.audio/) [中文](https://speech.fish.audio/zh/) [日本語](https://speech.fish.audio/ja/) [Portuguese (Brazil)](https://speech.fish.audio/pt/)
• I have searched for existing issues, including closed ones. [Search issues](https://github.com/fishaudio/fish-speech/issues)
• I confirm that I am using English to submit this report (我已阅读并同意 [Language Policy](Before you contribute, please read our 'Language Policy'. #515)).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:）
• Please do not modify this template and fill in all required fields.

Cloud or Self Hosted

Self Hosted (Source)

Environment Details

Environment

• Fish Speech: current development/test version
• GPU: NVIDIA GeForce RTX 5090 Laptop GPU
• PyTorch: 2.8.0+cu128
• CUDA: 12.8
• Python: 3.12
• OS: Windows 11 25H2, Build 26200.8875
• Backend: Fish Speech API server with --compile

Steps to Reproduce

1. Start the Fish Speech API server with --compile.
2. Provide a Punjabi speaker reference recording.
3. Use the following Punjabi sentence as the reference/input text:

ਅੱਜ ਮੌਸਮ ਬਹੁਤ ਵਧੀਆ ਹੈ, ਤੇ ਤੁਹਾਡੇ ਨਾਲ ਗੱਲ ਕਰਕੇ ਮੇਰਾ ਦਿਲ ਖੁਸ਼ ਹੋ ਗਿਆ।

4. Generate speech.
5. Generate the same sentence again using a Latin transliteration.
6. Compare the pronunciation, particularly words containing addak (ੱ) and Punjabi nasalization markers (ਂ / ੰ).

Comparison

Gurmukhi:

ਅੱਜ ਮੌਸਮ ਬਹੁਤ ਵਧੀਆ ਹੈ, ਤੇ ਤੁਹਾਡੇ ਨਾਲ ਗੱਲ ਕਰਕੇ ਮੇਰਾ ਦਿਲ ਖੁਸ਼ ਹੋ ਗਿਆ।

Plain Latin transliteration:

Ajj mausam bahut waddiya hai, te tuhade naal gal karke mera dil khush ho gya.

Diacritic-based Latin transliteration:

Ajj mausam bahut vadhīā hai, tē tuhāḍē nāl gall karke merā dil khuś hō giā.

The three versions produce noticeably different pronunciations.

✔️ Expected Behavior

Punjabi Gurmukhi input should preserve the pronunciation information represented by the Gurmukhi orthography.

In particular, addak (ੱ) should result in the appropriate consonant gemination, and Punjabi nasalization markers should contribute appropriately to pronunciation.

The pronunciation should also remain broadly consistent when the same linguistic content is represented in Gurmukhi and an unambiguous phonetic transliteration.

❌ Actual Behavior

The Gurmukhi version produces a noticeably different pronunciation from the Latin versions.

The difference is particularly noticeable in words containing Punjabi orthographic features such as addak and nasalization markers.

For example:

ਗੱਲ

contains addak and should represent consonant gemination. However, the generated pronunciation does not consistently sound equivalent to the corresponding explicitly geminated Latin representation:

gall

Interestingly, the diacritic-based Latin transliteration:

Ajj mausam bahut vadhīā hai, tē tuhāḍē nāl gall karke merā dil khuś hō giā.

produces substantially more natural Punjabi pronunciation in listening tests.

I generated the test sentence approximately ten times with the diacritic-based transliteration, with consistently good pronunciation.

This suggests that the issue may be related to Punjabi/Gurmukhi text normalization or grapheme-to-phoneme handling rather than the acoustic generation itself.

[avinashgola]

I investigated this against current main (befe400) and the official fishaudio/s2-pro tokenizer.

The tokenizer does not discard the reported Gurmukhi marks. Its backend uses NFC normalization, and these strings round-trip exactly through encode/decode:

• ਅੱਜ (includes addak U+0A71)
• ਗੱਲ (includes addak U+0A71)
• ਮਾਂ (includes bindi U+0A02)
• ਅੰਬ (includes tippi U+0A70)
• the complete sentence from this issue

I also traced native inference: ServeTTSRequest.text is passed unchanged from TTSInferenceEngine.send_Llama_request() into generate_long(). There is no Punjabi normalization, transliteration, or G2P step in that path. This matches the project description that S2-Pro handles multilingual text without language-specific preprocessing.

This points to a learned model/data pronunciation limitation rather than corruption of addak or nasalization markers in the source tokenizer.

A source-level workaround is possible: an explicit opt-in Gurmukhi-to-phonetic-Latin/ISO transliteration mode. I do not think it should silently rewrite Punjabi by default because transliteration introduces its own choices (including Punjabi schwa deletion), changes prompts across CLI/WebUI/server, and would need native-speaker audio validation.

Could a maintainer confirm which direction is wanted?

1. Treat this as a model/data quality issue, with a Punjabi pronunciation test set for a future checkpoint.
2. Accept an opt-in transliteration fallback in inference.

For either direction, fixed-seed before/after WAV samples plus the exact reference audio/text would make pronunciation validation reproducible. I can prepare deterministic preprocessing tests and the cross-interface wiring if option 2 is preferred.

[Whale-Dolphin]

Thanks for the detailed report and the side-by-side comparison!

We looked into it: the tokenizer and inference path pass Gurmukhi text through unchanged (addak, bindi and tippi all round-trip, as @avinashgola confirmed), so nothing is being dropped. The difference comes from the model itself handling Gurmukhi less well than a phonetically explicit Latin spelling.

The fix is therefore a Punjabi text-normalization front end, which is what #1321 proposes. Closing this as a duplicate so the discussion stays in one place — please follow #1321.